Learning in the Limit, General Topology and Modal Logicfitelson.org/bridges2/gierasimczuk.pdf · 2015-09-20 · Learning in the Limit, General Topology and Modal Logic Nina Gierasimczuk

Learning in the Limit, General Topologyand Modal Logic

Nina Gierasimczukresults obtained jointly with A. Baltag and S. Smets

Institute for Logic, Language and ComputationUniversity of Amsterdam

Bridges 2Rutgers University, September 19th, 2015

Outline

Introduction

Characterization of Learnability and Solvability

Constructive Order-driven Learning

Towards Epistemic Logic of Learnability

Intermediate (but Interesting ) Conclusions

Outline

Introduction

Characterization of Learnability and Solvability

Constructive Order-driven Learning

Towards Epistemic Logic of Learnability

Intermediate (but Interesting ) Conclusions

Background

I Learning and belief revision go their separate ways,

I conjecture dynamics is a common theme.

I What are the principles of this dynamics?

B B ′p

Background

I Learning and belief revision go their separate ways,

I conjecture dynamics is a common theme.

I What are the principles of this dynamics?

B B ′ B ′′p

Background

I Learning and belief revision go their separate ways,

I conjecture dynamics is a common theme.

I What are the principles of this dynamics?

B B ′ B ′′p

Truth-tracking!

Epistemic Spaces and Observables

DefinitionAn epistemic space is a pair S = (S ,O) consisting of a state space (a set ofpossible worlds) S and a countable set of observable properties O ⊆ P(S).

s t u w

UV

Learning: Streams of Observables

DefinitionLet S = (S ,O) be an epistemic space.

I A data stream is an infinite sequence ~O = (O0,O1, . . .) of data from O.

I A data sequence is a finite sequence σ = (σ0, . . . , σn).

DefinitionTake S = (S ,O) and s ∈ S . A data stream ~O is:

I sound with respect to s iff every element listed in ~O is true in s.

I complete with respect to s iff every observable true in s is listed in ~O.

We assume that data streams are sound and complete.

Learning: Learners and Conjectures

DefinitionLet S = (S ,O) be an epistemic space and let σ0, . . . , σn ∈ O. A learner is afunction L that on the input of S and data sequence (σ0, . . . , σn) outputs someset of worlds L(S, (σ0, . . . , σn)) ⊆ S , called a conjecture.

DefinitionS = (S ,O) is learnable by L if for every state s ∈ S we have that for every

sound and complete data stream ~O for s, there is n ∈ N s.t.:

L(S, (O0, . . . ,Ok)) = {s} for all k ≥ n.

An epistemic space S is learnable if it is learnable by a learner L.

Example of a Learnable Space

Let S = (S ,O) such that S = {sn | n ∈ N}, O = {pi | i ∈ N}, and for anyk ∈ N, pk = {si | 0 ≤ i ≤ k}. S is learnable.

p0 p1 p2 p3 p4. . .

s0 s2

s0 s1 s2 s3 s4

Example of a Learnable Space

Let S = (S ,O) such that S = {sn | n ∈ N}, O = {pi | i ∈ N}, and for anyk ∈ N, pk = {si | 0 ≤ i ≤ k}. S is learnable.

p0 p1 p2 p3 p4. . .

s0 s2

s0 s1 s2 s3 s4

Example of a Learnable Space

Let S = (S ,O) such that S = {sn | n ∈ N}, O = {pi | i ∈ N}, and for anyk ∈ N, pk = {si | 0 ≤ i ≤ k}. S is learnable.

p0 p1 p2 p3 p4. . .

s0

s2s0

s1 s2 s3 s4

Example of a Learnable Space

Let S = (S ,O) such that S = {sn | n ∈ N}, O = {pi | i ∈ N}, and for anyk ∈ N, pk = {si | 0 ≤ i ≤ k}. S is learnable.

p0 p1 p2 p3 p4. . .

s0

s2s0 s1

s2

s3 s4

Example of a Non-Learnable Space

Consider S = (S ,O), where S := {sn | n ∈ N} ∪ {s∞}, and O = {pi | i ∈ N},and for any k ∈ N, pk := {sk , sk+1, . . .} ∪ {s∞}. S is not learnable.

p0 p1 p2 p3. . .

s1 s2

s0 s1 s2 s3 s∞

Example of a Non-Learnable Space

Consider S = (S ,O), where S := {sn | n ∈ N} ∪ {s∞}, and O = {pi | i ∈ N},and for any k ∈ N, pk := {sk , sk+1, . . .} ∪ {s∞}. S is not learnable.

p0 p1 p2 p3. . .

s1 s2

s0 s1 s2 s3 s∞

Example of a Non-Learnable Space

Consider S = (S ,O), where S := {sn | n ∈ N} ∪ {s∞}, and O = {pi | i ∈ N},and for any k ∈ N, pk := {sk , sk+1, . . .} ∪ {s∞}. S is not learnable.

p0 p1 p2 p3. . .

s1

s2

s0

s1

s2 s3 s∞

Example of a Non-Learnable Space

Consider S = (S ,O), where S := {sn | n ∈ N} ∪ {s∞}, and O = {pi | i ∈ N},and for any k ∈ N, pk := {sk , sk+1, . . .} ∪ {s∞}. S is not learnable.

p0 p1 p2 p3. . .

s1 s2

s0 s1 s2 s3 s∞

Example of a Non-Learnable Space

Consider S = (S ,O), where S := {sn | n ∈ N} ∪ {s∞}, and O = {pi | i ∈ N},and for any k ∈ N, pk := {sk , sk+1, . . .} ∪ {s∞}. S is not learnable.

p0 p1 p2 p3. . .

s1

s2s0 s1

s2

s3 s∞

Questions, Answers, and Problems

DefinitionA question Q is a partition of S , whose cells Ai are called answers to Q.Given s ∈ A ⊆ S , A ∈ Q is called the answer to Q at s, denoted As .

DefinitionQ′ is a refinement of Q if all answers of Q is a disjoint union of answers of Q′.

DefinitionA problem P is a pair (S,Q) consisting of S = (S ,O) and Q over S .P′ = (S,Q′) is a refinement of P = (S,Q) if Q′ is a refinement of Q.

Illustration

t s

u v

VU

PQ

Illustration

t s

u v

VU

PQ

Solving in the Limit

DefinitionA learning method L solves a problem P = (S,Q) in the limit iff for every state

s ∈ S and every data stream ~O for s, there exists some k ∈ N such that:

L(S, ~O[n]) ⊆ As for all n ≥ k.

A problem is solvable in the limit if there is a learner that solves it in the limit.

General Topology

DefinitionA topology τ over a set S is a collection of subsets of S (open sets) s.t.:

1. ∅ ∈ τ ,

2. S ∈ τ ,

3. for any X ⊆ τ ,⋃

X ∈ τ , and

4. for any finite X ⊆ τ we have⋂

X ∈ τ.

DefinitionTake a set X ⊆ S .

1. The interior of X : Int(X ) =⋃{U ∈ τ | U ⊆ X}.

2. A subset Y ⊆ S is closed if an only if its complement, Y c is open.

3. The closure of X : X = (Int(X c))c =⋂{Y | X ⊆ Y and Y is closed}.

Separability by observations: Illustration

s t u w

UV

(a) t and u are not separable

t s

VU

(b) weakly separated space, T0

V

t

U

s

(c) strongly separated space, T1

Locally Closed and Constructible Sets

DefinitionA topology τ is Td iff for every s ∈ S there is a U ∈ τ such that U \ {s} ∈ τ ,i.e., for every s ∈ S there is a U ∈ τ such that {s} = U ∩ {s}.Td is a separation property between T0 and T1.

DefinitionA set A is locally closed if A = U ∩ C , where U is open and C is closed.

A set is constructible if it is a finite disjoint union of locally closet sets.

An ω-constructible set is a countable union of locally closed sets.

Outline

Introduction

Characterization of Learnability and Solvability

Constructive Order-driven Learning

Towards Epistemic Logic of Learnability

Intermediate (but Interesting ) Conclusions

The Topology Associated with an Epistemic Space

DefinitionThe topology τS associated with an epistemic space S = (S ,O) is a collectionof subsets of S of the following properties:

1. for any O ∈ O it is the case that O ∈ τS2. ∅ ∈ τS,

3. S ∈ τS,

4. for any U ⊆ τS,⋃

U ∈ τS, and

5. for any x , y ∈ τS we have x ∩ y ∈ τS.

s t u w

UV

The Topology Associated with an Epistemic Space

DefinitionThe topology τS associated with an epistemic space S = (S ,O) is a collectionof subsets of S of the following properties:

1. for any O ∈ O it is the case that O ∈ τS2. ∅ ∈ τS,

3. S ∈ τS,

4. for any U ⊆ τS,⋃

U ∈ τS, and

5. for any x , y ∈ τS we have x ∩ y ∈ τS.

s t u w

UV

Characterization of Solvability in the Limit

TheoremA problem P = (S,Q) is solvable in the limit iff Q has a locally closedrefinement.

CorollaryAn epistemic space S = (S ,O) is learnable in the limit iff it satisfies the Td

separation axiom.

Outline

Introduction

Characterization of Learnability and Solvability

Constructive Order-driven Learning

Towards Epistemic Logic of Learnability

Intermediate (but Interesting ) Conclusions

Order-driven learning: Motivation

I Belief Revision: minimal states give beliefs.

I Computational Learning Theory: co-learning, learning by erasing.

I Philosophy of Science: Ockham’s razor.

Conditioning

DefinitionConditioning wrt a prior ≤ on S , is defined in the following way:

L≤(O1, . . . ,On) := Min≤

(n⋂

i=1

Oi

)

whenever⋂

i Oi has any minimal elements; and otherwise:

L≤(O1, . . . ,On) :=n⋂

i=1

Oi .

DefinitionConditioning is said to be standard if the prior ≤ is well-founded.

TheoremNon-standard conditioning is a universal problem solving method.

Outline

Introduction

Characterization of Learnability and Solvability

Constructive Order-driven Learning

Towards Epistemic Logic of Learnability

Intermediate (but Interesting ) Conclusions

Logic for Learnability

Since learnability is about potentially successful changes of beliefsone expects some doxastic logic to capture it and to reason about it.

Relational semantics for modal logic

Definition (Syntax)

Take countable set of propositional symbols P.

ϕ := p | ¬ϕ | ϕ ∧ ϕ | �ϕ,

for all p ∈ P, the usual abbreviations are ∨, →, and ♦.

Definition (Semantics)

Given a model M = (W ,R, v), where v : P → ℘(W ), and a state x ∈W :

M, x |= p iff x ∈ v(p) for each p ∈ PM, x |= ¬ϕ iff not M, x |= ϕM, x |= ϕ ∧ ψ iff M, x |= ϕ and M, x |= ψM, x |= �ϕ iff for all y ∈W : if xRy then M, y |= ϕand dually:M, x |= ♦ϕ iff there is y ∈W : xRy and M, y |= ϕ

Relational semantics for modal logic

Definition (Syntax)

Take countable set of propositional symbols P.

ϕ := p | ¬ϕ | ϕ ∧ ϕ | �ϕ,

for all p ∈ P, the usual abbreviations are ∨, →, and ♦.

Definition (Semantics)

Given a model M = (W ,R, v), where v : P → ℘(W ), and a state x ∈W :

M, x |= p iff x ∈ v(p) for each p ∈ PM, x |= ¬ϕ iff not M, x |= ϕM, x |= ϕ ∧ ψ iff M, x |= ϕ and M, x |= ψM, x |= �ϕ iff for all y ∈W : if xRy then M, y |= ϕand dually:M, x |= ♦ϕ iff there is y ∈W : xRy and M, y |= ϕ

Some Axioms and Their Epistemic Meaning

Rules

(MP) if ` ϕ and ` ϕ→ ψ, then ` ψ(N) if ` ϕ, then ` �ϕ

Axioms

(K) �(ϕ→ ψ)→ (�ϕ→ �ψ) (omniscience)

(T) �ϕ→ ϕ (truthfullness/reflexivity)

(D) �ϕ→ ¬�¬ϕ (consistency/seriality)

(4) �ϕ→ ��ϕ (positive introspection/transitivity)

(5) ¬�ϕ→ �¬�ϕ (negative introspection/Euclidean-ness)

Ax is a logic of a class of models M iff Ax is sound and complete wrt M.

Some Axioms and Their Epistemic Meaning

Rules

(MP) if ` ϕ and ` ϕ→ ψ, then ` ψ(N) if ` ϕ, then ` �ϕ

Axioms

(K) �(ϕ→ ψ)→ (�ϕ→ �ψ) (omniscience)

(T) �ϕ→ ϕ (truthfullness/reflexivity)

(D) �ϕ→ ¬�¬ϕ (consistency/seriality)

(4) �ϕ→ ��ϕ (positive introspection/transitivity)

(5) ¬�ϕ→ �¬�ϕ (negative introspection/Euclidean-ness)

Ax is a logic of a class of models M iff Ax is sound and complete wrt M.

Can we use modal logic on topologies?Relational � vs Topological � := Int

�ϕ �ϕ

DefinitionLet P be a set of propositional symbols. A topological model (or atopo-model) M = (X ,O, v) is a topological space τ = (X ,O) together with avaluation function v : P → ℘(X ).

Can we use modal logic on topologies?Relational � vs Topological � := Int

�ϕ �ϕ

DefinitionLet P be a set of propositional symbols. A topological model (or atopo-model) M = (X ,O, v) is a topological space τ = (X ,O) together with avaluation function v : P → ℘(X ).

Topological Topo-semantics for Modal Logic

DefinitionTruth of modal formulas is defined inductively at points x in a topo-modelM = (X ,O, v) in the following way:

M, x |= p iff x ∈ v(p) for each p ∈ PM, x |= ¬ϕ iff not M, x |= ϕM, x |= ϕ ∧ ψ iff M, x |= ϕ and M, x |= ψM, x |= �ϕ iff there is U ∈ τ(x ∈ U and for all y ∈ U: M, y |= ϕ)

and dually:M, x |= ♦ϕ iff for all U ∈ τ(x ∈ U → there is y ∈ U: M, y |= ϕ)

Sound and Complete Topo-Axiomatizations

Rules

(MP) if ` ϕ and ` ϕ→ ψ, then ` ψ(N) if ` ϕ, then ` �ϕ

Axioms

(K) �(ϕ→ ψ)→ (�ϕ→ �ψ)

(T) �ϕ→ ϕ

(D) �ϕ→ ¬�¬ϕ(4) �ϕ→ ��ϕ

(5) ¬�ϕ→ �¬�ϕ

S4=Topo

S4 is the topo-logic of all topological spaces (McKinsey & Tarski 1944).

Sound and Complete Topo-Axiomatizations

Rules

(MP) if ` ϕ and ` ϕ→ ψ, then ` ψ(N) if ` ϕ, then ` �ϕ

Axioms

(K) �(ϕ→ ψ)→ (�ϕ→ �ψ)

(T) �ϕ→ ϕ

(D) �ϕ→ ¬�¬ϕ

(4) �ϕ→ ��ϕ

(5) ¬�ϕ→ �¬�ϕ

S4=Topo

S4 is the topo-logic of all topological spaces (McKinsey & Tarski 1944).

Sound and Complete Topo-Axiomatizations

Rules

(MP) if ` ϕ and ` ϕ→ ψ, then ` ψ(N) if ` ϕ, then ` �ϕ

Axioms

(K) �(ϕ→ ψ)→ (�ϕ→ �ψ)

(T) �ϕ→ ϕ

(D) �ϕ→ ¬�¬ϕ

(4) �ϕ→ ��ϕ

(5) ¬�ϕ→ �¬�ϕ

S4=Topo

S4 is the topo-logic of all topological spaces (McKinsey & Tarski 1944).

What about Td -spaces (the learning spaces)?

Td is not topo-definable.

Learnable spaces are not topo-definable.

Luckily, we can once again change the way we view �.

�ϕ �ϕ

Topological d-semantics

DefinitionTruth of modal formulas is defined inductively at points x in a topo-modelM = (X , τ, v) in the following way:

M, x |=d p iff x ∈ v(p) for each p ∈ PM, x |=d ¬ϕ iff not M, x |=d ϕM, x |=d ϕ ∧ ψ iff M, x |=d ϕ and M, x |=d ψM, x |=d �ϕ iff ∃U ∈ τ(x ∈ U & ∀y ∈ U − {x} M, y |=d ϕ)

and dually:

M, x |=d ♦ϕ iff ∀U ∈ τ(x ∈ U → ∃y ∈ U − {x} M, y |=d ϕ)

Sound and Complete d-Axiomatizations

Rules

(MP) if ` ϕ and ` ϕ→ ψ, then ` ψ(N) if ` ϕ, then ` �ϕ

Axioms

(K) �(ϕ→ ψ)→ (�ϕ→ �ψ)

(4) �ϕ→ ��ϕ

K4 is the d-logic of all Td -spaces.

K4=T

d

KD45 Doxastic d-logic (Steinvold 2006)

Because independent reasons (e.g., Stalnaker) one may want B:=� to be:

(K) �(ϕ→ ψ)→ (�ϕ→ �ψ)

(D) �ϕ→ ¬�¬ϕ(4) �ϕ→ ��ϕ

(5) ¬�ϕ→ �¬�ϕ

Theorem (Steinsvold 2006)

KD45 is a sound and complete d-axiomatization of DSO spaces.

DSO stands for ‘derived sets are open’. DSO are Td -spaces (by 4), in which allderived sets are open (5), except that there are no open singletons (D).

Questions

But DSO ⊂ Td .

So what do we talk about when we talk about beliefs in learning?

Should conjectures be interpreted as beliefs?

What if one restricts conjectures to only those which are ‘proper’ beliefs?

Outline

Introduction

Characterization of Learnability and Solvability

Constructive Order-driven Learning

Towards Epistemic Logic of Learnability

Intermediate (but Interesting ) Conclusions

Conclusions

I Topological characterization of learnability & solvability in the limit.

I Universality of conditioning as a problem solving method.

I Use of stratification-like topological techniques.

Moreover:

I Learnable spaces are Td .

I Td -spaces are not topo-definable.

I Learnability is not topo-definable.

I Learnability cannot be expressed by solely topo-definable belief operators.

I The existing topo- and d-logics of belief are to fluffy to capture learnability.

THANK YOU!

A. Baltag, N. Gierasimczuk, and S. Smets. On the solvability of inductive problems: A

study in epistemic topology. Proceedings of TARK’15 (ILLC PP-2015-13), 2015.

A. Baltag, N. Gierasimczuk, and S. Smets. Truth tracking by belief revision. ILLC

PP-2014-20 (to appear in Studia Logica), 2015.

A. Baltag, N. Gierasimczuk, and S. Smets. Belief revision as a truth-tracking process. In

K. Apt, editor, Proceedings of TARK’11, pages 187-190. ACM, 2011.

A. Baltag, N. Bezhanishvili, A. Ozgun, and S. Smets. The Topology of Belief, Belief Revision

and Defeasible Knowledge. Proceedings of the LORI’13, 2013

M. de Brecht and A. Yamamoto. Topological properties of concept spaces. Information and

Computation, 208(4):327340, 2010.

N. Gierasimczuk. Knowing Ones Limits. Logical Analysis of Inductive Inference. PhD thesis,

Universiteit van Amsterdam, The Netherlands, 2010.

N. Gierasimczuk, D. de Jongh, and V. F. Hendricks. Logic and learning. In A. Baltag and S.

Smets, editors, Johan van Benthem on Logical and Informational Dynamics. Springer, 2014.

THANK YOU!

A. Baltag, N. Gierasimczuk, and S. Smets. On the solvability of inductive problems: A

study in epistemic topology. Proceedings of TARK’15 (ILLC PP-2015-13), 2015.

A. Baltag, N. Gierasimczuk, and S. Smets. Truth tracking by belief revision. ILLC

PP-2014-20 (to appear in Studia Logica), 2015.

A. Baltag, N. Gierasimczuk, and S. Smets. Belief revision as a truth-tracking process. In

K. Apt, editor, Proceedings of TARK’11, pages 187-190. ACM, 2011.

A. Baltag, N. Bezhanishvili, A. Ozgun, and S. Smets. The Topology of Belief, Belief Revision

and Defeasible Knowledge. Proceedings of the LORI’13, 2013

M. de Brecht and A. Yamamoto. Topological properties of concept spaces. Information and

Computation, 208(4):327340, 2010.

N. Gierasimczuk. Knowing Ones Limits. Logical Analysis of Inductive Inference. PhD thesis,

Universiteit van Amsterdam, The Netherlands, 2010.

N. Gierasimczuk, D. de Jongh, and V. F. Hendricks. Logic and learning. In A. Baltag and S.

Smets, editors, Johan van Benthem on Logical and Informational Dynamics. Springer, 2014.